Professional Documents
Culture Documents
Anyway, for some reason I got into my head that with the
4A transformer it should be possible to recognize speech.
You just take the speech data, transform it into the
frequency domain. Match it up against patterns for words.
And presto, magic, HAL would be born. And this idea
always stuck around in the back of my mind.
Peter Brown: Well, sort of. They were pre trained, like
ChatGPT is. They're generative like ChatGPT is. And like
with ChatGPT, the entire goal was to take a bunch of text
and predict the next word of text. But remember, back then
it was before the internet and before the NVIDIA gaming
chips. So, we had a lot less data and a lot less computer
power. And we were constantly scrounging around for data
from places like the Canadian Parliament, the Federal
Register, the American Reading House for the Blind. We
even got data from the depositions in the IBM anti-trust
lawsuit.
"What do you mean?" "I don't know," said the man. "Is it,"
he asked, said the man. "They are not not to be a good
idea. The first time, I was a good idea. She was a good
idea." "Certainly," I said. "What's the matter?" "May I be
able to get the money," said the man. "Scott was a good
idea." "Mrs. King," Nick said, "I don't know what I mean.
Take a look at the door. He was a good idea." "I don't know
what I mean. Didn't you," he said.
Now, I don't know where all this "I don't know what I mean"
is coming from. And I don't know where this "good idea"
stuff is coming from. Maybe that's the way they talked in
these IBM depositions. They constantly talked about good
ideas. I don't know. Anyway, you can see there's been a lot
of progress in the past 35 years.
Peter Brown: Well, back then the party line was that to
have a computer do tasks that humans do, you had to feed
the computer with a bunch of domain-specific knowledge.
So, for example, if you wanted to do French to English
translation, you had to tell the computer about French
grammar.
Peter Brown: Bob Mercer and I had each consulted for IDA
in Princeton. And Jim Simons used to work at IDA. And he
had hired an IDA cryptographer named Nick Paterson. And
Nick recommended us to Jim. And so, Jim then sent us
some letters.
Now, I had no idea who Jim was. But I learned that his
original partner was someone named Lenny Baum who had
invented the EM algorithm. And we knew all about the EM
algorithm because we used it heavily in our work on
machine translation and in speech recognition. So, I
figured if Lenny was Jim's partner, Jim had to be a serious
customer.
Then three things happened. First, Bob had a second
daughter accepted to Stanford. But he couldn't afford to
pay for her to go to Stanford on his IBM salary. So, she had
to go to the agricultural school at Cornell, which offered
scholarships to New York State residents.
So, at the end of 2002, Bob and I also took over the rest of
the technical side of the firm, which included the trading of
currencies, bonds, options, and futures. Now, our plan was
to use the equities code that we and others had developed
to trade these other instruments. But we recognized they
would not be so great for morale to tell, say, one of the
futures researchers, "You know all that code you spent the
last decade of your life developing, guess what, we're going
to throw it out." So, we had to spend quite a bit of time
getting everyone to buy into our plan.
So, that was the genesis of the institutional funds. The goal
was to create products with attractive, risk adjusted
returns.
Then we took a big loss one day. So, I worked through the
night trying to understand what was going on. The next
day we took another big loss. And I, again, worked through
that night. So, now it's the third day and I hadn't slept for,
I don't know, 48 or 50 hours. And I was sitting in a meeting
with Jim and a few others when the head of production
knocked on the door and asked to speak with me. I walked
out of the meeting, and he told me we were down again by
a large amount.
A couple days later I went into Jim's office and told him
that I'd screwed up in not appreciating the risk we were
taking and said that if he wanted me to resign, I would
resign. But he responded, "Peter, quite the opposite. Now
that you've been through such a stressful losing period,
you're far more valuable to me and to the firm than you
were before." Now, that response really tells you something
about Jim Simons.
So, I called into work when it was going on and I got Kim,
Jim's assistant. And she said, "Jim wants you to get back
here as soon as you're physically able." So, I raced out. I
found a taxi, leaving my family to fend for themselves at
Newark Airport. And pushed the driver to drive as fast as
he could from Newark to Long Island. I ran into my office,
and I found Jim, Bob, Paul Broder, who was head of risk
control, all holed up. And the office was full of cigarette
smoke. I could barely breathe. And then there was this, I
remember seeing this, 16 oz cup full of Jim's cigarette
butts. And I'm thinking, like, why do they have to do this in
my office?
And they were all staring through the haze at the computer
screens trying to figure out what was going on. And Jim
was interpreting every little wiggle and various graphs. He
was really scared. And he wanted to cut back and hard.
Paul also wanted to cut back.
Raj, I'm sure you know, the head of risk control always
wants to cut back.
But I looked at the data and saw that the model had these
enormous predictions, the likes of which I had never seen
before. It was clear to me what was going on. People were
dumping positions that were correlated with their own
positions. And they were driving prices to ridiculous levels.
I felt they had to come back. I argued that we should not
cut back. That this was going to be the greatest money-
making opportunity we'd ever seen. And if anything, we
should increase our positions.
For me, the most important takeaway from the flash crash
was that exchanges needed to increase their capacity to
handle spikes in volume. They've done this. And since
then, things have been better.
Raj Mahajan: It's what you say after the yes that's
actually more interesting. Is it true that you once gave
someone a raise so that you could call him in the middle of
the night?
Speaker: All price references and market forecasts correspond to the date of
this recording. This podcast should not be copied, distributed, published, or
reproduced in whole, or in part. The information contained in this podcast does
not constitute research or recommendation from any Goldman Sachs entity to
the listener.
Neither Goldman Sachs, nor any of its affiliates, makes any representation or
warranty as to the accuracy or completeness of the statements or any
information contained in this podcast and any liability, therefore, including in
respect of direct, indirect or consequential loss or damage is expressly
disclaimed.
The views expressed in this podcast are not necessarily those of Goldman
Sachs. And Goldman Sachs is not providing any financial, economic, legal,
accounting, or tax advice or recommendations in this podcast. In addition, the
receipt of this podcast by any listener is not to be taken as constituting the
giving of investment advice by Goldman Sachs to that listener, nor to constitute
such person, a client of any Goldman Sachs entity.